The randomized algorithms used in Data Streaming applications, at least in the papers I have read are Monte Carlo algorithms. This is because they need not always produce a correct solution even allowing for approximation bounds. There is always a possibility epsilon of an error. Hence the Yao's minimax conjecture does not apply to these. Hence it is consistent, that, while deterministic algorithms produce a linear lower bound on space requirements these algorithms can indeed give a sublinear bound on space.